Anthropic Just Reserved Enough Computing Power to Rival a Small City
Anthropic locked in 3.5 gigawatts of computing power — city-scale — through a deal with Google and Broadcom. Behind the huge number is a simple story about why AI companies are racing to buy chips years before they exist.
Anthropic — the company behind the Claude AI assistant — just signed a deal with Google and Broadcom for about 3.5 gigawatts of computing power. That number is hard to picture, so here’s a yardstick: 3.5 gigawatts is roughly the electricity a mid-sized European city uses. That’s how much raw computing an AI company now considers a reasonable shopping list.
Let’s unpack what’s actually being bought. Training and running big AI models requires enormous banks of specialized chips, humming away in warehouse-sized data centers that draw serious electricity. In this deal, Google Cloud provides much of the muscle, and Broadcom will manufacture custom chips designed specifically for Anthropic’s needs. Most of it sits in the US, and it’s meant to power Claude’s training from 2027 onward — meaning the company is buying capacity years before it’s switched on.
Why commit so far ahead? Demand. Anthropic says its revenue is running at about $30 billion a year, up from roughly $9 billion at the end of 2025, and the number of large companies spending over $1 million a year on Claude passed 1,000 — and doubled in under two months. When demand grows that fast, the risk isn’t overbuying; it’s running out of capacity and turning customers away. There’s also a strategic angle: nearly everyone’s chips come from one supplier (NVIDIA), and waiting in that line is a bottleneck. By having Google and Broadcom build a dedicated chip supply, Anthropic reduces its dependence on that queue.
Step back and the pattern is striking. The AI industry has shifted from renting computing by the hour to signing multi-year, multi-billion-dollar contracts for hardware that hasn’t been made yet. It’s a giant bet that demand for AI will keep outrunning the world’s ability to supply it. That bet might be right or wildly optimistic — nobody knows — but the companies making it are effectively pre-paying for their spot at the front of the line.
What this means for you: If you’re a business choosing an AI provider, deals like this are a signal of staying power — a company with years of guaranteed capacity is less likely to hit the “sorry, we’re full” walls that occasionally slowed model access in the past. For everyone else, it’s a useful reality check: the AI in your apps runs on genuinely massive, expensive infrastructure. That’s also why efficiency matters — and why a smaller model that runs on your own machine, sipping ordinary household power, is quietly appealing for tasks that don’t need a city’s worth of chips.
Sources
Source: https://www.anthropic.com/news/google-broadcom-partnership-compute
Can an AI Be an Inventor? Japan's Top Court Says No — Joining the Rest of the World
Japan's Supreme Court settled a question that's been fought over worldwide: only humans can be named as inventors on a patent. Simple enough — but it leaves a much trickier question wide open for anyone using AI in their work.
